Upload 5 files

configuration_friday.py

@@ -0,0 +1,5 @@
+from .modeling_friday import FridayConfig
+
+__all__ = [
+    "FridayConfig",
+]

configuration_phi3.py

@@ -0,0 +1,226 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Phi-3 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Phi3Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-3
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the
+    [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32064):
+            Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Phi3Model`].
+        hidden_size (`int`, *optional*, defaults to 3072):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 8192):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        resid_pdrop (`float`, *optional*, defaults to 0.0):
+            Dropout probability for mlp outputs.
+        embd_pdrop (`int`, *optional*, defaults to 0.0):
+            The dropout ratio for the embeddings.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio after computing the attention scores.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model might ever be used with.
+        original_max_position_embeddings (`int`, *optional*, defaults to 4096):
+            The maximum sequence length that this model was trained with. This is used to determine the size of the
+            original RoPE embeddings when using long scaling.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon value used for the RMSNorm.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`dict`, *optional*):
+            The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must
+            contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be `longrope` and
+            the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size
+            divided by the number of attention heads divided by 2.
+        partial_rotary_factor (`float`, *optional*, defaults to 1.0):
+            Percentage of the query and keys which will have rotary embedding. Must be between 0.0 and 1.0.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            The id of the "beginning-of-sequence" token.
+        eos_token_id (`int`, *optional*, defaults to 32000):
+            The id of the "end-of-sequence" token.
+        pad_token_id (`int`, *optional*, defaults to 32000):
+            The id of the padding token.
+        sliding_window (`int`, *optional*):
+            Sliding window attention window size. If `None`, no sliding window is applied.
+
+    Example:
+
+    ```python
+    >>> from transformers import Phi3Model, Phi3Config
+
+    >>> # Initializing a Phi-3 style configuration
+    >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
+
+    >>> # Initializing a model from the configuration
+    >>> model = Phi3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "phi3"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=32064,
+        hidden_size=3072,
+        intermediate_size=8192,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        resid_pdrop=0.0,
+        embd_pdrop=0.0,
+        attention_dropout=0.0,
+        hidden_act="silu",
+        max_position_embeddings=4096,
+        original_max_position_embeddings=4096,
+        initializer_range=0.02,
+        rms_norm_eps=1e-5,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        partial_rotary_factor=1.0,
+        bos_token_id=1,
+        eos_token_id=32000,
+        pad_token_id=32000,
+        sliding_window=None,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attention_dropout = attention_dropout
+        self.hidden_act = hidden_act
+        self.max_position_embeddings = max_position_embeddings
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.partial_rotary_factor = partial_rotary_factor
+        self._rope_scaling_adjustment()
+        self._rope_scaling_validation()
+        self.sliding_window = sliding_window
+
+        super().__init__(
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            pad_token_id=pad_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    def _rope_scaling_adjustment(self):
+        """
+        Adjust the `type` of the `rope_scaling` configuration for backward compatibility.
+        """
+        if self.rope_scaling is None:
+            return
+
+        rope_scaling_type = self.rope_scaling.get("type", None)
+
+        # For backward compatibility if previous version used "su" or "yarn"
+        if rope_scaling_type is not None and rope_scaling_type in ["su", "yarn"]:
+            self.rope_scaling["type"] = "longrope"
+
+    def _rope_scaling_validation(self):
+        """
+        Validate the `rope_scaling` configuration.
+        """
+        if self.rope_scaling is None:
+            return
+
+        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
+            raise ValueError(
+                "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "
+                f"got {self.rope_scaling}"
+            )
+        rope_scaling_type = self.rope_scaling.get("type", None)
+        rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)
+        rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)
+        if rope_scaling_type is None or rope_scaling_type not in ["longrope"]:
+            raise ValueError(f"`rope_scaling`'s type field must be one of ['longrope'], got {rope_scaling_type}")
+        if not (
+            isinstance(rope_scaling_short_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
+            )
+        rotary_ndims = int(self.hidden_size // self.num_attention_heads * self.partial_rotary_factor)
+        if not len(rope_scaling_short_factor) == rotary_ndims // 2:
+            raise ValueError(
+                f"`rope_scaling`'s short_factor field must have length {rotary_ndims // 2}, got {len(rope_scaling_short_factor)}"
+            )
+        if not (
+            isinstance(rope_scaling_long_factor, list)
+            and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
+        ):
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
+            )
+        if not len(rope_scaling_long_factor) == rotary_ndims // 2:
+            raise ValueError(
+                f"`rope_scaling`'s long_factor field must have length {rotary_ndims // 2}, got {len(rope_scaling_long_factor)}"
+            )

friday_map.py

@@ -0,0 +1,4 @@
+from .configuration_friday import FridayConfig        # your real class
+from .modeling_friday import FridayForCausalLM
+
+__all__ = ["FridayForCausalLM", "FridayConfig"]

modeling_friday.py

@@ -0,0 +1,1248 @@
+from __future__ import annotations
+
+# Model Constants
+IMAGE_TOKEN = "<image>"
+IMG_START_TOKEN   = "<img_start>"
+IMG_END_TOKEN     = "<img_end>"
+IGNORE_INDEX      = -100
+PAD_FOR_EOS       = -300
+
+
+
+
+
+import torch
+import torch.nn.functional as F
+
+from PIL import Image
+
+
+import torch
+
+def mask_token_segment(
+               start_id: int,
+               end_id: int,
+               input_ids: torch.Tensor,
+               fill_value: int = -100):
+    """
+    Replace *every* token from each `start_id` **through** its matching `end_id`
+    (boundaries included) with `fill_value`.  Any spans that start with some
+    other token are left untouched.
+
+    Works on CUDA, TorchScript, batched via vmap, etc.—no Python loops.
+    """
+    if input_ids.dim() != 1:
+        raise ValueError("`input_ids` must be 1-D")
+
+    device = input_ids.device
+    n       = input_ids.size(0)
+
+    # where the *target* start-tokens and end-tokens sit
+    start_pos = (input_ids == start_id).nonzero(as_tuple=True)[0]      # ascending
+    end_pos   = (input_ids == end_id).nonzero(as_tuple=True)[0]        # ascending
+
+    if start_pos.numel() == 0:
+        return input_ids.clone()
+
+    # ── pair every start with the first end that comes *after* it ────────────────
+    # searchsorted gives the insertion index into the (sorted) end positions
+    idx_in_end = torch.searchsorted(end_pos, start_pos, right=False)
+
+    have_match = idx_in_end < end_pos.size(0)                # safety: drop unmatched
+    start_pos  = start_pos[have_match]
+    end_pos    = end_pos[idx_in_end[have_match]]
+
+    # (rare) guard against pathological orderings
+    keep = end_pos > start_pos
+    start_pos, end_pos = start_pos[keep], end_pos[keep]
+
+    if start_pos.numel() == 0:
+        return input_ids
+
+    # ── differential “scan-line” trick to build the span mask in O(N) ───────────
+    # +1 at each start index, -1 at the element *after* each end
+    delta = torch.zeros(n + 1, dtype=torch.int8, device=device)
+    delta[start_pos]        += 1
+    delta[end_pos + 1]      -= 1          # +1 is safe because delta is length n+1
+
+    inside = torch.cumsum(delta[:-1], dim=0) > 0   # boolean mask, incl. boundaries
+
+    # ── apply ────────────────────────────────────────────────────────────────────
+    out = input_ids.clone()
+    out[inside] = fill_value
+    return out
+
+
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                print(name, 'no ignore status')
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+# Borrowed from peft.util.get_peft_model_state_dict
+def get_peft_state_maybe_zero_3(named_params, bias):
+    if bias == "none":
+        to_return = {k: t for k, t in named_params if "lora_" in k}
+    elif bias == "all":
+        to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
+    elif bias == "lora_only":
+        to_return = {}
+        maybe_lora_bias = {}
+        lora_bias_names = set()
+        for k, t in named_params:
+            if "lora_" in k:
+                to_return[k] = t
+                bias_name = k.split("lora_")[0] + "bias"
+                lora_bias_names.add(bias_name)
+            elif "bias" in k:
+                maybe_lora_bias[k] = t
+        for k, t in maybe_lora_bias:
+            if bias_name in lora_bias_names:
+                to_return[bias_name] = t
+    else:
+        raise NotImplementedError
+    to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()}
+    return to_return
+
+
+def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
+    to_return = {k: t for k, t in named_params if "lora_" not in k}
+    if require_grad_only:
+        to_return = {k: t for k, t in to_return.items() if t.requires_grad}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def find_all_linear_names(modules):
+    lora_module_names = set()
+    for name, module in modules():
+        if isinstance(module, torch.nn.Linear):
+            names = name.split('.')
+            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
+
+    if 'lm_head' in lora_module_names:  # needed for 16-bit
+        lora_module_names.remove('lm_head')
+    return list(lora_module_names)
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+def pad_and_stack(img_list, pad_value=0.0):
+    """
+    img_list : list[Tensor]  each (C, H, W) already *normalised*
+    pad_value: float or tuple/list of 3 floats (one per channel)
+               Use 0.0 if your processor has already centred to mean 0.
+    Returns
+    -------
+    batch : Tensor  (B, C, H_max, W_max)
+    """
+
+    # 1. target square size ---------------------------------------------------
+    h_max = max(t.shape[1] for t in img_list)
+    w_max = max(t.shape[2] for t in img_list)
+    H, W  = max(h_max, w_max), max(h_max, w_max)
+
+    # 2. create padded copies -------------------------------------------------
+    padded = []
+    for img in img_list:
+        c, h, w = img.shape
+        canvas   = img.new_full((c, H, W), pad_value)     # filled with mean/zeros
+        canvas[:, :h, :w] = img                    # top-left corner
+        padded.append(canvas)
+
+    return torch.stack(padded, 0)                  # (B,C,H,W)
+
+
+
+
+
+#  ------------------------------------------------------------------------------------------
+#  Copyright (c) 2024 Baifeng Shi.
+#  All rights reserved.
+#
+#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
+#  ------------------------------------------------------------------------------------------
+
+import torch
+
+def split_chessboard(x, num_split):
+    """
+        x: b * c * h * w
+        Deividing x into num_split**2 sub-squares, and concatenate all the sub-squares on the batch dimension
+    """
+    B, C, H, W = x.shape
+    assert H % num_split == 0 and W % num_split == 0
+    h, w = H // num_split, W // num_split
+    x_split = torch.cat([x[:, :, i*h:(i+1)*h, j*w:(j+1)*w] for i in range(num_split) for j in range(num_split)], dim=0)
+    return x_split
+
+def merge_chessboard(x, num_split):
+    """
+        x: b * c * h * w
+        Assuming x contains num_split**2 sub-squares concatenated along batch dimension, merge the sub-squares back to the original whole square.
+        (inverse of split_chessboard)
+    """
+    B, C, H, W = x.shape
+    assert B % (num_split**2) == 0
+    b = B // (num_split**2)
+    x_merge = torch.cat([torch.cat([x[(i*num_split + j)*b:(i*num_split + j + 1)*b] for j in range(num_split)], dim=-1)
+                         for i in range(num_split)], dim=-2)
+    return x_merge
+
+def batched_forward(model, x, batch_size=-1):
+    if batch_size == -1:
+        return model(x)
+    else:
+        x_batched = x.split(batch_size)
+        outs = [model(x) for x in x_batched]
+        return torch.cat(outs, dim=0)
+
+
+
+
+
+
+
+#  ------------------------------------------------------------------------------------------
+#  Copyright (c) 2024 Baifeng Shi.
+#  All rights reserved.
+#
+#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
+#  ------------------------------------------------------------------------------------------
+
+import math
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+
+def multiscale_forward(model, input, scales=None, img_sizes=None, max_split_size=None, resize_output_to_idx=0, num_prefix_token=0,
+            output_shape='bnc', split_forward=False):
+
+    # print(f"Input shape: {input.shape}")
+
+    assert input.dim() == 4, "Input image must be in the shape of BxCxHxW."
+    assert input.shape[2] == input.shape[3], "Currently only square images are supported."
+    assert output_shape in ['bnc', 'bchw'], "Output shape should be either BxNxC (e.g., ViT) or BxCxHxW (e.g., ConvNet)."
+    assert output_shape == 'bnc' or num_prefix_token == 0, "For ConvNet there shouldn't be any prefix token."
+
+    b, c, input_size, _ = input.shape
+
+    # image size for each scale
+    assert scales is not None or img_sizes is not None, "Please assign either scales or img_sizes."
+    img_sizes = img_sizes or [int(input_size * scale) for scale in scales]
+
+    # prepare multiscale inputs
+    max_split_size = max_split_size or input_size   # The maximum size of each split of image. Set as the input size by default
+    num_splits = [math.ceil(size / max_split_size) for size in img_sizes]   # number of splits each scale
+    input_multiscale = []
+    for size, num_split in zip(img_sizes, num_splits):
+        x = F.interpolate(input.to(torch.float32), size=size, mode='bicubic').to(input.dtype)
+        x = split_chessboard(x, num_split=num_split)
+        input_multiscale.append(x)
+
+    # run feedforward on each scale
+    outs_multiscale = [batched_forward(model, x, b) if split_forward else model(x) for x in input_multiscale]
+    if num_prefix_token > 0:
+        outs_prefix_multiscale = [out[:, :num_prefix_token] for out in outs_multiscale]
+        outs_multiscale = [out[:, num_prefix_token:] for out in outs_multiscale]
+    if output_shape == 'bnc':
+        outs_multiscale = [rearrange(out, 'b (h w) c -> b c h w', h=int(out.shape[1] ** 0.5), w=int(out.shape[1] ** 0.5))
+                           for out in outs_multiscale]
+
+    # merge outputs of different splits for each scale separately
+    outs_multiscale = [merge_chessboard(out, num_split=num_split) for num_split, out in zip(num_splits, outs_multiscale)]
+
+    # interpolate outputs from different scales and concat together
+    output_size = outs_multiscale[resize_output_to_idx].shape[-2]
+    out = torch.cat([F.interpolate(outs_multiscale[i].to(torch.float32), size=output_size,
+                                   mode='area').to(outs_multiscale[i].dtype)
+                     for i in range(len(outs_multiscale))], dim=1)
+    if output_shape == 'bnc':
+        out = rearrange(out, 'b c h w -> b (h w) c')
+    if num_prefix_token > 0:
+        # take the mean of prefix tokens from different splits for each scale
+        outs_prefix_multiscale = [torch.stack(out.split(b, dim=0), dim=0).mean(dim=0) for out in outs_prefix_multiscale]
+        out_prefix_multiscale = torch.cat(outs_prefix_multiscale, dim=-1)
+        out = torch.cat([out_prefix_multiscale, out], dim=1)
+
+    return out
+
+
+
+
+
+import torch
+import torch.nn as nn
+
+class MLPAdapter(nn.Module):
+
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers=2, activation='gelu', checkpoint_path=None, device=None, **kwargs):
+        """
+        Initialize the MLPAdapter with the given dimensions and activation function.
+
+        Args:
+            input_dim (int): Input dimension.
+            hidden_dim (int): Hidden dimension.
+            output_dim (int): Output dimension.
+            layers (int): Number of layers in the MLP.
+            activation (str): Activation function to use ('gelu' or 'relu').
+        """
+        super().__init__()
+        self.num_layers = num_layers
+        self.activation = activation
+        self.output_dim = output_dim
+
+        # Define the first layer
+        layers_list = [nn.Linear(input_dim, hidden_dim, device=device)]
+        if activation == 'gelu':
+            layers_list.append(nn.GELU())
+        elif activation == 'relu':
+            layers_list.append(nn.ReLU())
+        else:
+            raise ValueError("Unsupported activation function. Use 'gelu' or 'relu'.")
+        
+        # Define the subsequent layers
+        for _ in range(1, num_layers):
+            layers_list.append(nn.Linear(hidden_dim, hidden_dim, device=device))
+            if activation == 'gelu':
+                layers_list.append(nn.GELU())
+            elif activation == 'relu':
+                layers_list.append(nn.ReLU())
+        
+        # Define the final output layer
+        layers_list.append(nn.Linear(hidden_dim, output_dim, device=device))
+        self.mlp = nn.Sequential(*layers_list)
+
+        # Load checkpoint if provided
+        if checkpoint_path:
+            self.load_state_dict(torch.load(checkpoint_path, map_location=device), strict=False)
+            print(f"Loaded MLPAdapter from {checkpoint_path}")
+        
+        if device:
+            self.to(device)
+
+    def forward(self, x):
+        """
+        Forward pass through the MLPAdapter.
+
+        Args:
+            x (torch.Tensor): Input tensor.
+
+        Returns:
+            torch.Tensor: Output tensor after passing through the MLP.
+        """
+        return self.mlp(x)
+    
+
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import PIL.Image
+from typing import List
+
+from transformers import AutoModel, AutoImageProcessor
+
+
+class FastVitVisionTower(nn.Module):
+    def __init__(self, pretrained_model_name_or_path, model_params={}, pad_to_square=True, **kwargs):
+        super().__init__()
+
+        self.is_loaded = False
+        self.pretrained_model_name_or_path = pretrained_model_name_or_path
+        self.model_params = model_params
+        self.pad_to_square = pad_to_square
+        self.load_model()
+
+    @property
+    def output_dim(self):
+        return self.vision_tower.config.embed_dim if self.vision_tower else None
+    
+    def load_model(self):
+        if self.is_loaded:
+            return
+        self.image_processor = AutoImageProcessor.from_pretrained(self.pretrained_model_name_or_path)
+        self.image_processor.crop_size = self.image_processor.size
+        self.vision_tower = AutoModel.from_pretrained(
+            self.pretrained_model_name_or_path,
+            **self.model_params,
+        )
+        self.vision_tower.requires_grad_(False)
+
+        self.is_loaded = True
+    
+    def preprocess_images(self, imgs: List[PIL.Image.Image], pad_and_stack_tensors=True) -> torch.Tensor:
+        img_mean = tuple(int(x * 255) for x in self.image_processor.image_mean)
+        if self.pad_to_square:
+            imgs = [expand2square(img, img_mean) for img in imgs]
+        
+        imgs = [self.image_processor(img, do_resize=True, do_center_crop=False, return_tensors="pt")['pixel_values'][0] for img in imgs]
+        
+
+        if pad_and_stack_tensors:
+            imgs = pad_and_stack(imgs, pad_value=0.0)
+            imgs = imgs.to(dtype=torch.float32, device=self.device)
+        
+        return imgs
+
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_feature = self.vision_tower(
+                    image.to(device=self.device, dtype=self.dtype).unsqueeze(0)
+                )
+                image_features.append(image_feature)
+        else:
+            image_features = self.vision_tower(
+                images.to(device=self.device, dtype=self.dtype),
+            )
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.embed_dim, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_tower.dtype
+
+    @property
+    def device(self):
+        return self.vision_tower.device
+
+    @property
+    def config(self):
+        if self.is_loaded:
+            return self.vision_tower.config
+        else:
+            return self.cfg_only
+
+    @property
+    def hidden_size(self):
+        return self.config.embed_dim
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2
+
+
+class FastVitVisionTowerS2(FastVitVisionTower):
+    def __init__(self, pretrained_model_name_or_path, s2_scales, model_params={}, **kwargs):
+        self.s2_scales = list(map(int, s2_scales.split(',')))
+        self.s2_scales.sort()
+        self.s2_split_size = self.s2_scales[0]
+        self.s2_image_size = self.s2_scales[-1]
+
+        super().__init__(pretrained_model_name_or_path, model_params)
+
+        self.multiscale_forward = multiscale_forward
+    
+    @property
+    def output_dim(self):
+        return (2*self.vision_tower.config.embed_dim) if self.vision_tower else None
+
+    def load_model(self):
+        if self.is_loaded:
+            return
+        
+        super().load_model()
+        self.image_processor.size = self.image_processor.crop_size = {
+            "height": self.s2_image_size,
+            "width":  self.s2_image_size
+        }
+
+    def forward_feature(self, images):
+        image_size = self.vision_tower.config.image_size
+        if images.shape[2] != image_size or images.shape[3] != image_size:
+            images = F.interpolate(
+                images,
+                size=(image_size, image_size),
+                mode="bilinear",
+                align_corners=False,
+                antialias=True
+            )    
+
+        return self.vision_tower(
+            images.to(device=self.device, dtype=self.dtype),
+        )
+
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_feature = self.multiscale_forward(
+                    self.forward_feature, 
+                    image.unsqueeze(0),
+                    img_sizes=self.s2_scales, 
+                    max_split_size=self.s2_split_size
+                )
+                image_features.append(image_feature)
+        else:
+            image_features = self.multiscale_forward(
+                self.forward_feature, 
+                images, 
+                img_sizes=self.s2_scales,
+                max_split_size=self.s2_split_size
+            )
+
+        return image_features
+
+    @property
+    def hidden_size(self):
+        return self.config.embed_dim * len(self.s2_scales)
+
+
+
+
+
+import torch
+import torch.nn as nn
+
+import PIL.Image
+from typing import List
+
+from transformers import SiglipVisionModel, SiglipImageProcessor, SiglipVisionConfig
+
+
+class SiglipVisionTower(nn.Module):
+    def __init__(self, pretrained_model_name_or_path, model_params={}, pad_to_square=True, **kwargs):
+        super().__init__()
+
+        self.is_loaded = False
+        self.pretrained_model_name_or_path = pretrained_model_name_or_path
+        self.model_params = model_params
+        self.pad_to_square = pad_to_square
+        self.select_layer = -2
+        self.load_model()
+
+    @property
+    def output_dim(self):
+        return self.vision_tower.config.hidden_size if self.vision_tower else None
+    
+    def load_model(self):
+        if self.is_loaded:
+            return
+        self.image_processor = SiglipImageProcessor.from_pretrained(self.pretrained_model_name_or_path)
+        self.image_processor.crop_size = self.image_processor.size
+        self.vision_tower = SiglipVisionModel.from_pretrained(
+            self.pretrained_model_name_or_path,
+            **self.model_params,
+        )
+        self.vision_tower.requires_grad_(False)
+
+        self.is_loaded = True
+    
+    def preprocess_images(self, imgs: List[PIL.Image.Image], pad_and_stack_tensors=True) -> torch.Tensor:
+        img_mean = tuple(int(x * 255) for x in self.image_processor.image_mean)
+        if self.pad_to_square:
+            imgs = [expand2square(img, img_mean) for img in imgs]
+        imgs = [self.image_processor(img, return_tensors="pt")['pixel_values'][0] for img in imgs]
+
+        if pad_and_stack_tensors:
+            imgs = pad_and_stack(imgs, pad_value=0.0)
+            imgs = imgs.to(dtype=torch.float32, device=self.device)
+        
+        return imgs
+
+    def feature_select(self, image_forward_outs):
+        image_features = image_forward_outs.hidden_states[self.select_layer]
+
+        return image_features
+
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0),
+                                                      output_hidden_states=True)
+                image_feature = self.feature_select(image_forward_out).to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype),
+                                                   output_hidden_states=True)
+            image_features = self.feature_select(image_forward_outs).to(images.dtype)
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_tower.dtype
+
+    @property
+    def device(self):
+        return self.vision_tower.device
+
+    @property
+    def config(self):
+        if self.is_loaded:
+            return self.vision_tower.config
+        else:
+            return self.cfg_only
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2
+
+
+class SiglipVisionTowerS2(SiglipVisionTower):
+    def __init__(self, pretrained_model_name_or_path, s2_scales, model_params={}, **kwargs):
+        self.s2_scales = list(map(int, s2_scales.split(',')))
+        self.s2_scales.sort()
+        self.s2_split_size = self.s2_scales[0]
+        self.s2_image_size = self.s2_scales[-1]
+
+        super().__init__(pretrained_model_name_or_path, model_params)
+
+        self.multiscale_forward = multiscale_forward
+
+        self.image_processor.size['height'] = self.image_processor.size['width'] = self.s2_image_size
+        self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size
+    
+    @property
+    def output_dim(self):
+        return (2*self.vision_tower.config.hidden_size) if self.vision_tower else None
+
+    def load_model(self):
+        if self.is_loaded:
+            return
+        
+        super().load_model()
+        self.image_processor.size['height'] = self.image_processor.size['width'] = self.s2_image_size
+        self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size
+
+    def forward_feature(self, images):
+        image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype),
+                                               output_hidden_states=True)
+        image_features = self.feature_select(image_forward_outs).to(images.dtype)
+        return image_features
+
+    def forward(self, images):
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_feature = self.multiscale_forward(
+                    self.forward_feature, 
+                    image.unsqueeze(0),
+                    img_sizes=self.s2_scales, 
+                    max_split_size=self.s2_split_size
+                )
+                image_features.append(image_feature)
+        else:
+            image_features = self.multiscale_forward(
+                self.forward_feature, 
+                images, 
+                img_sizes=self.s2_scales,
+                max_split_size=self.s2_split_size
+            )
+
+        return image_features
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size * len(self.s2_scales)
+
+
+
+
+
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision import transforms
+
+from typing import List, Tuple, Optional, Union
+
+import PIL
+
+from transformers import AutoTokenizer, AutoConfig
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from .configuration_phi3 import Phi3Config
+from .modeling_phi3 import Phi3Model, Phi3ForCausalLM
+
+
+DEFAULT_CFG_SPECIAL_TOKENS = {
+    "image_token_id": 200029,
+    "image_start_token_id": 200030,
+    "image_end_token_id": 200031,
+}
+DEFAULT_CFG_VISION_TOWER = {
+    "pretrained_model_name_or_path": "kevin510/fast-vit-hd",
+    "type": "fastvit",
+    "s2_scales": "512,1024",
+    "use_s2": True,
+    "pad_to_square": True,
+    "freeze": False,
+    "model_params": { "trust_remote_code": True }
+}
+DEFAULT_CFG_VISION_ADAPTER = {
+    "input_dim": 6144,
+    "hidden_dim": 3072,
+    "output_dim": 3072,
+    "layers": 2,
+    "activation": "gelu",
+    "freeze": False,
+}
+
+
+class FridayConfig(Phi3Config):
+    model_type = "friday"
+
+    def __init__(self, 
+            base_model_name_or_path: str | None = "microsoft/Phi-4-mini-reasoning",
+            delay_load=False, 
+            tokenizer_model_max_length=None,
+            **kwargs
+        ):
+        base_kwargs = {}
+        if base_model_name_or_path is not None:
+            base_cfg = AutoConfig.from_pretrained(
+                base_model_name_or_path,
+                trust_remote_code=True,   # Phi‑4 uses custom code in the repo
+            )
+            base_kwargs = base_cfg.to_dict()
+
+        merged = {**base_kwargs, **kwargs}
+        self.delay_load = delay_load
+        self.tokenizer_model_max_length = tokenizer_model_max_length
+
+        self._cfg_vision_tower = DEFAULT_CFG_VISION_TOWER.copy()
+        if "cfg_vision_tower" in kwargs:
+            self._cfg_vision_tower.update(kwargs["cfg_vision_tower"])
+
+        self._cfg_vision_adapter = DEFAULT_CFG_VISION_ADAPTER.copy()
+        if "cfg_vision_adapter" in kwargs:
+            self._cfg_vision_adapter.update(kwargs["cfg_vision_adapter"])
+
+        self._cfg_special_tokens = DEFAULT_CFG_SPECIAL_TOKENS.copy()
+        if "cfg_special_tokens" in kwargs:
+            self._cfg_special_tokens.update(kwargs["cfg_special_tokens"])
+
+        super().__init__(**merged)
+        
+    
+    @property
+    def cfg_vision_tower(self):
+        return self._cfg_vision_tower
+
+    @cfg_vision_tower.setter
+    def cfg_vision_tower(self, value):
+        if not value:
+            raise ValueError("Name cannot be empty")
+        self._cfg_vision_tower.update(value)
+    
+
+    @property
+    def cfg_vision_adapter(self):
+        return self._cfg_vision_adapter
+    
+    @cfg_vision_adapter.setter
+    def cfg_vision_adapter(self, value):
+        if not value:
+            raise ValueError("Name cannot be empty")
+        self._cfg_vision_adapter.update(value)
+    
+    @property
+    def cfg_special_tokens(self):
+        return self._cfg_special_tokens
+    
+    @cfg_special_tokens.setter
+    def cfg_special_tokens(self, value):
+        if not value:
+            raise ValueError("Name cannot be empty")
+        self._cfg_special_tokens.update(value)
+
+
+class FridayModel(Phi3Model):
+    config_class = FridayConfig
+    
+    def __init__(self, config: FridayConfig):
+        super().__init__(config)
+
+        self.cfg_vision_adapter = config.cfg_vision_adapter
+        self.cfg_vision_tower = config.cfg_vision_tower
+
+        self.vision_tower = None
+        self.mm_projector    = None
+        if not config.delay_load:
+            self.initialize_vision_modules()
+    
+    def get_vision_tower(self):
+        return self.vision_tower
+    
+    def initialize_vision_modules(self):
+        if self.vision_tower is not None:
+            return
+
+        if self.cfg_vision_tower.get("type", "siglip").lower() == "siglip":
+            if self.cfg_vision_tower.get("use_s2", True):
+                self.vision_tower = SiglipVisionTowerS2(**self.cfg_vision_tower)
+            else:
+                self.vision_tower = SiglipVisionTower(**self.cfg_vision_tower)
+        elif self.cfg_vision_tower.get("type", "siglip").lower() == "fastvit":
+            if self.cfg_vision_tower.get("use_s2", True):
+                self.vision_tower = FastVitVisionTowerS2(**self.cfg_vision_tower)
+            else:
+                self.vision_tower = FastVitVisionTower(**self.cfg_vision_tower)
+        else:
+            raise ValueError(f"Unsupported vision tower type: {self.cfg_vision_tower.get('type', 'siglip')}. Supported types are 'siglip' and 'fastvit'.")
+        
+        self.vision_tower.load_model()
+        self.mm_projector = MLPAdapter(**self.cfg_vision_adapter)
+
+        if self.cfg_vision_tower.get("freeze", False):
+            self.set_vision_tower_requires_grad(False)
+        
+        if self.cfg_vision_adapter.get("freeze", False):
+            self.set_vision_adapter_requires_grad(False)
+    
+    def compute_image_features(self, imgs: torch.Tensor) -> torch.Tensor:
+        features = self.vision_tower(imgs)
+        if isinstance(features, list):
+            features = torch.stack(features, dim=1)
+        return self.mm_projector(features)
+    
+    def set_vision_tower_requires_grad(self, requires_grad: bool):
+        if self.vision_tower is not None:
+            for param in self.vision_tower.parameters():
+                param.requires_grad = requires_grad
+        else:
+            raise ValueError("Vision tower is not initialized. Please call initialize_vision_modules() first.")
+    
+    def set_vision_adapter_requires_grad(self, requires_grad: bool):
+        if self.mm_projector is not None:
+            for param in self.mm_projector.parameters():
+                param.requires_grad = requires_grad
+        else:
+            raise ValueError("Vision adapter is not initialized. Please call initialize_vision_modules() first.")
+    
+    def set_vision_tower_dtype(self, dtype: torch.dtype):
+        if self.vision_tower is not None:
+            for p in self.vision_tower.parameters():
+                p.data = p.data.to(dtype)
+        else:
+            raise ValueError("Vision tower is not initialized. Please call initialize_vision_modules() first.")
+    
+    def set_vision_adapter_dtype(self, dtype: torch.dtype):
+        if self.mm_projector is not None:
+            for p in self.mm_projector.parameters():
+                p.data = p.data.to(dtype)
+        else:
+            raise ValueError("Vision adapter is not initialized. Please call initialize_vision_modules() first.")
+    
+    def is_vision_tower_frozen(self):
+        if self.vision_tower is not None:
+            return all(not p.requires_grad for p in self.vision_tower.parameters())
+        else:
+            raise ValueError("Vision tower is not initialized. Please call initialize_vision_modules() first.")
+    
+    def is_vision_adapter_frozen(self):
+        if self.mm_projector is not None:
+            return all(not p.requires_grad for p in self.mm_projector.parameters())
+        else:
+            raise ValueError("Vision adapter is not initialized. Please call initialize_vision_modules() first.")
+
+
+class FridayForCausalLM(Phi3ForCausalLM):
+    config_class = FridayConfig
+
+    def __init__(self, config: FridayConfig):
+        super().__init__(config)
+
+        self.config = config
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.image_token_id = config.cfg_special_tokens["image_token_id"]
+        self.image_start_id       = config.cfg_special_tokens["image_start_token_id"]
+        self.image_end_id         = config.cfg_special_tokens["image_end_token_id"]
+
+        self.model = FridayModel(config)
+        self.post_init()
+    
+    def get_model(self) -> FridayModel:
+        return self.model
+    
+    def get_vision_tower(self) -> SiglipVisionTower:
+        return self.model.get_vision_tower()
+    
+    def get_vision_adapter(self) -> MLPAdapter:
+        return self.model.mm_projector
+
+    def get_llm_parameters(self, exclude_lora: bool = False):
+        return [
+            p for n, p in self.named_parameters() 
+            if "vision_tower" not in n and "mm_projector" not in n and (not exclude_lora or ("lora_" not in n))
+        ]
+
+    def get_llm_named_modules(self):
+        return {n: m for n, m in self.named_modules() if "vision_tower" not in n and "mm_projector" not in n}
+    
+    def set_llm_requires_grad(self, requires_grad: bool, exclude_lora: bool = True):
+        for n, p in self.named_parameters():
+            if exclude_lora and ("lora_A" in n or "lora_B" in n):
+                continue
+            if "vision_tower" in n or "mm_projector" in n:
+                continue
+            p.requires_grad = requires_grad
+    
+    def set_vision_tower_requires_grad(self, requires_grad: bool):
+        self.model.set_vision_tower_requires_grad(requires_grad)
+
+    def set_vision_adapter_requires_grad(self, requires_grad: bool):
+        self.model.set_vision_adapter_requires_grad(requires_grad)
+    
+    def set_llm_dtype(self, dtype: torch.dtype):
+        for p in self.get_llm_parameters():
+            p.data = p.data.to(dtype)
+
+    def set_vision_tower_dtype(self, dtype: torch.dtype):
+        self.model.set_vision_tower_dtype(dtype)
+    
+    def set_vision_adapter_dtype(self, dtype: torch.dtype):
+        self.model.set_vision_adapter_dtype(dtype)
+    
+    def is_llm_frozen(self):
+        return all(not p.requires_grad for p in self.get_llm_parameters())
+    
+    def is_vision_tower_frozen(self):
+        return self.model.is_vision_tower_frozen()
+    
+    def is_vision_adapter_frozen(self):
+        return self.model.is_vision_adapter_frozen()
+    
+    
+    
+    def initialize_vision_modules(self):
+        self.model.initialize_vision_modules()
+    
+    def get_multimodal_input_embeddings(self, input_ids, image_features, return_labels=True) -> torch.Tensor:
+        emb_start_image_id = self.model.embed_tokens(torch.tensor([self.image_start_id], device=self.device))
+        emb_end_image_id   = self.model.embed_tokens(torch.tensor([self.image_end_id], device=self.device))
+        id_ignore = torch.tensor([IGNORE_INDEX], device=self.device)
+
+        # repetition‑penalty safety ????
+        # input_ids[input_ids == self.image_token_id] = 0
+
+        
+        # Iterate over each batch item
+        embeds_list, labels_list = [], []
+        for batch_id, item_ids in enumerate(input_ids):
+            
+            image_token_positions = (item_ids == self.image_token_id).nonzero(as_tuple=True)[0]
+            if len(image_token_positions) != image_features[batch_id].shape[0]:
+                raise ValueError(
+                    f"Mismatch between number of image tokens ({len(image_token_positions)}) and number of image features ({image_features[batch_id].shape[0]})"
+                )
+
+
+            cursor = 0
+            emb_parts, lbl_parts = [], []
+            for indx_image, image_token_pos in enumerate(image_token_positions):
+                if image_token_pos > cursor:
+                    span = item_ids[cursor:image_token_pos]
+                    emb_parts.append(self.model.embed_tokens(span))
+                    lbl_parts.append(span)
+
+                # <image_start>
+                emb_parts.append(emb_start_image_id)
+                lbl_parts.append(id_ignore)
+
+                # vision embeddings
+                image_tokens = image_features[batch_id][indx_image]
+                if image_tokens.shape[0] == 1 and image_tokens.ndim == 3:
+                    image_tokens = image_tokens.squeeze(0)
+                emb_parts.append(image_tokens)
+                lbl_parts.append(id_ignore.repeat(image_tokens.shape[0]))
+
+                # <image_end>
+                emb_parts.append(emb_end_image_id)
+                lbl_parts.append(id_ignore)
+
+                cursor = image_token_pos + 1
+            
+            # tail text
+            if cursor < item_ids.shape[0]:
+                tail = item_ids[cursor:]
+                emb_parts.append(self.model.embed_tokens(tail))
+                lbl_parts.append(tail)
+            
+            embeds_list.append(torch.cat(emb_parts, dim=0))
+            labels_list.append(torch.cat(lbl_parts, dim=0))
+    
+        return (embeds_list, labels_list) if return_labels else embeds_list
+
+    def prepare_inputs_for_multimodal(
+        self,
+        input_ids: torch.LongTensor,
+        images: List[List[PIL.Image.Image]], # B x N
+        position_ids: Optional[torch.LongTensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_values: Optional[List[torch.FloatTensor]],
+        labels: Optional[torch.LongTensor],
+    ) -> Tuple[Optional[torch.Tensor], Optional[torch.LongTensor], Optional[torch.Tensor], Optional[List[torch.FloatTensor]], torch.Tensor, Optional[torch.Tensor]]:
+        
+        # ─────────────────── early return (no image / streaming step) ───────────────────
+        # if we have already processed images and are in a streaming step we can skip the multimodal processing
+        # but we need to ensure the attention mask and position ids are correct
+
+        if past_key_values is not None and attention_mask is not None and input_ids.shape[1] == 1:
+            tgt = past_key_values[-1][-1].shape[-2] + 1
+            attention_mask = torch.cat(
+                [attention_mask,
+                torch.ones((attention_mask.size(0),
+                            tgt - attention_mask.size(1)),
+                            dtype=attention_mask.dtype,
+                            device=attention_mask.device)],
+                dim=1,
+            )
+            position_ids = (attention_mask.sum(dim=1, keepdim=True) - 1).long()
+
+            return input_ids, position_ids, attention_mask, past_key_values, None, labels
+
+        # ─────────────────────────── images: (B, N) ───────────────────────────
+        if isinstance(images, list) and isinstance(images[0], list):
+            # images is a list of lists, each containing multiple images, B x N
+            # e.g. [[img1, img2], [img3, img4]]
+            assert len(images) == input_ids.shape[0], f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}"
+            image_features = []
+            for sublst_images in images:
+                if len(sublst_images) == 0:
+                    image_features.append(torch.zeros((0, self.get_model().mm_projector.output_dim), device=self.device))
+                else:
+                    if isinstance(sublst_images[0], PIL.Image.Image):
+                        image_features.append(
+                            self.model.compute_image_features(
+                                self.model.vision_tower.preprocess_images(sublst_images, pad_and_stack_tensors=True)
+                            )
+                        )
+                    elif isinstance(sublst_images[0], torch.Tensor):
+                        # This should be a list of tensors of pre-processed images, [(N X 3 X W x H), ...]
+                        image_features.append(
+                            self.model.compute_image_features(sublst_images)
+                        )
+        elif isinstance(images, list) and isinstance(images[0], PIL.Image.Image):
+            # images is a list of images for a single batch item, 1 x N
+            # e.g. [img1, img2, img3]
+            assert input_ids.shape[0] == 1, f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}"
+            image_features = [
+                self.model.compute_image_features(
+                    self.model.vision_tower.preprocess_images(images, pad_and_stack_tensors=True)
+                )
+            ]
+        elif isinstance(images, list) and isinstance(images[0], torch.Tensor):
+            # This should be a list of tensors of pre-processed images, [(N X 3 X W x H), ...]
+            # The list length should match the batch size
+            assert input_ids.shape[0] == len(images), f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}"
+            image_features = [
+                self.model.compute_image_features(imgs) for imgs in images
+            ]
+        elif isinstance(images, PIL.Image.Image):
+            # images is a single image, 1 x 1
+            # e.g. img1
+            assert input_ids.shape[0] == 1, f"Batch size mismatch: {len(images)} vs {input_ids.shape[0]}"
+            image_features = [
+                self.model.compute_image_features(
+                    self.model.vision_tower.preprocess_images([images])
+                )
+            ]
+        else:
+            raise ValueError(f"Unsupported images format: {type(images)}. Expected list of PIL images, a single PIL image or a Tensor of pre-processed images")
+        
+        # ─────────────────────────── image_features: (B x N x D) ───────────────────────────
+        if isinstance(image_features, list):
+            assert input_ids.shape[0] == len(image_features), f"Incorrectly formatted image_features: list length should match batch size"
+            assert isinstance(image_features[0], torch.Tensor), f"Incorrectly formatted image_features: list items should be tensors"
+        elif isinstance(image_features, torch.Tensor):
+            assert input_ids.shape[0] == image_features.shape[0], f"Incorrectly formatted image_features: tensor should match batch size"
+        
+
+        # ───────────────────────────── pad handling prelims ──────────────────────────────
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
+        else:
+            attention_mask = attention_mask.bool()
+        if position_ids is None:
+            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
+
+        input_ids_nopad = [ids[mask] for ids, mask in zip(input_ids, attention_mask)]
+        embeds_list, labels_list = self.get_multimodal_input_embeddings(
+            input_ids_nopad,
+            image_features,
+            return_labels=True
+        )
+
+        # ───────────────────── truncate then pad back to rectangle ──────────────────────
+        new_input_embeds = torch.nn.utils.rnn.pad_sequence(
+            embeds_list,
+            batch_first=True,
+            padding_value=0.0
+        ).to(dtype=self.dtype)
+
+        new_labels = torch.nn.utils.rnn.pad_sequence(
+            labels_list,
+            batch_first=True,
+            padding_value=IGNORE_INDEX
+        ).long()
+
+        if self.config.tokenizer_model_max_length is not None:
+            new_input_embeds = new_input_embeds[:, :self.config.tokenizer_model_max_length]
+            new_labels       = new_labels[:, :self.config.tokenizer_model_max_length]
+
+        
+        
+
+        # ────────────────────────────── attention mask and position ids ────────────────
+        
+        attention_mask = (
+            torch.arange(new_input_embeds.size(1), device=input_ids.device)
+                  .unsqueeze(0)
+            < torch.tensor([e.size(0) for e in embeds_list],
+                           device=input_ids.device).unsqueeze(1)
+        )
+
+        raw_pos = attention_mask.cumsum(dim=1) - 1
+        position_ids = raw_pos.masked_fill(~attention_mask, 0).long()
+
+        if not self.training:
+            new_labels = None
+
+        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
+
+    
+    
+    # ------------------------------------------------------------------
+    def forward(
+            self,
+            input_ids: torch.LongTensor = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+            inputs_embeds: Optional[torch.FloatTensor] = None,
+            labels: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+            cache_position: Optional[torch.LongTensor] = None,
+            logits_to_keep: Union[int, torch.Tensor] = 0,
+            images: Optional[PIL.Image.Image] = None,
+            **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+
+        is_multi_modal = images is not None and not (
+            (
+                isinstance(images, list) and (len(images) == 0 or all(i == [] for i in images))
+            )
+        )
+
+
+        if inputs_embeds is None and is_multi_modal:
+            (
+                input_ids,
+                position_ids,
+                attention_mask,
+                past_key_values,
+                inputs_embeds,
+                labels
+            ) = self.prepare_inputs_for_multimodal(
+                input_ids=input_ids,
+                images=images,
+                position_ids=position_ids,
+                attention_mask=attention_mask,
+                past_key_values=past_key_values,
+                labels=labels,
+            )
+
+            if cache_position is not None and inputs_embeds is not None and cache_position.shape[0] != inputs_embeds.shape[1]:
+                cache_position = torch.arange(inputs_embeds.shape[1], device=self.device)
+        
+        
+        return Phi3ForCausalLM.forward(
+            self,
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            labels=labels,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            logits_to_keep=logits_to_keep,
+            **kwargs
+        )
+    
+    def print_device_configuration(self):
+        print("*************Device Configuration*********")
+        if len(self.get_llm_parameters()) > 0:
+            llm_device = set({str(p.device) for p in self.get_llm_parameters()})
+            llm_dtype = set({p.dtype for p in self.get_llm_parameters()})
+            print(f"LLM Parameters:\t\t\tdevice: {llm_device}\tdtype: {llm_dtype}\tfrozen: {self.is_llm_frozen()}")
+        else:
+            print("LLM parameters have not been initialized")
+        
+        if self.get_model().vision_tower is not None:
+            vt_device = set({str(p.device) for p in self.get_model().vision_tower.parameters()})
+            vt_dtype = set({p.dtype for p in self.get_model().vision_tower.parameters()})
+            print(f"Vision Tower Parameters:\tdevice: {vt_device}\tdtype: {vt_dtype}\tfrozen: {self.is_vision_tower_frozen()}")
+        else:
+            print("Vision tower parameters have not been initialized")
+
+        if self.get_model().mm_projector is not None:
+            mm_device = set({str(p.device) for p in self.get_model().mm_projector.parameters()})
+            mm_dtype = set({p.dtype for p in self.get_model().mm_projector.parameters()})
+            print(f"MM Projector Parameters:\tdevice: {mm_device}\tdtype: {mm_dtype}\tfrozen: {self.is_vision_adapter_frozen()}")
+        else:
+            print("MM Projector parameters have not been initialized")
+        print("******************************************")
+
+
+
+def build_tokenizer(base_model_id: str) -> Tuple[AutoTokenizer, dict]:
+    tok = AutoTokenizer.from_pretrained(base_model_id, padding_side="right")
+    specials = {t: tok.convert_tokens_to_ids(t) for t in [IMAGE_TOKEN, IMG_START_TOKEN, IMG_END_TOKEN] if t in tok.vocab}
+    if len(specials) < 3:
+        n = tok.add_tokens([IMAGE_TOKEN, IMG_START_TOKEN, IMG_END_TOKEN], special_tokens=True)
+        tok.pad_token = tok.eos_token
+        specials = {
+            "image": tok.convert_tokens_to_ids(IMAGE_TOKEN),
+            "start": tok.convert_tokens_to_ids(IMG_START_TOKEN),
+            "end": tok.convert_tokens_to_ids(IMG_END_TOKEN),
+        }
+    return tok, specials
+

modeling_phi3.py

@@ -0,0 +1,1181 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""PyTorch Phi-3 model."""
+
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    TokenClassifierOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    LossKwargs,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+from .configuration_phi3 import Phi3Config
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "microsoft/Phi-3-mini-4k-instruct"
+_CONFIG_FOR_DOC = "Phi3Config"
+
+
+class Phi3MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+
+        self.config = config
+        self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
+        self.activation_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+        up_states = self.gate_up_proj(hidden_states)
+
+        gate, up_states = up_states.chunk(2, dim=-1)
+        up_states = up_states * self.activation_fn(gate)
+
+        return self.down_proj(up_states)
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    rotary_dim = cos.shape[-1]
+    q_rot, q_pass = q[..., :rotary_dim], q[..., rotary_dim:]
+    k_rot, k_pass = k[..., :rotary_dim], k[..., rotary_dim:]
+
+    q_embed = torch.cat([(q_rot * cos) + (rotate_half(q_rot) * sin), q_pass], dim=-1)
+    k_embed = torch.cat([(k_rot * cos) + (rotate_half(k_rot) * sin), k_pass], dim=-1)
+    return q_embed, k_embed
+
+
+class Phi3Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        op_size = config.num_attention_heads * self.head_dim + 2 * (config.num_key_value_heads * self.head_dim)
+        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
+        self.qkv_proj = nn.Linear(config.hidden_size, op_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        qkv = self.qkv_proj(hidden_states)
+        query_pos = self.config.num_attention_heads * self.head_dim
+        query_states = qkv[..., :query_pos]
+        key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
+        value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
+
+        query_states = query_states.view(hidden_shape).transpose(1, 2)
+        key_states = key_states.view(hidden_shape).transpose(1, 2)
+        value_states = value_states.view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
+                logger.warning_once(
+                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
+                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+                )
+            else:
+                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=getattr(self.config, "sliding_window", None),
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class Phi3RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Phi3RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class Phi3DecoderLayer(nn.Module):
+    def __init__(self, config: Phi3Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = Phi3Attention(config=config, layer_idx=layer_idx)
+        self.mlp = Phi3MLP(config)
+        self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.config = config
+        self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
+        self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`):
+                input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+                Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
+                `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+            past_key_value (`Cache`, *optional*): cached past key and value projection states
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + self.resid_attn_dropout(hidden_states)  # main diff with Llama
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + self.resid_mlp_dropout(hidden_states)  # main diff with Llama
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class Phi3RotaryEmbedding(nn.Module):
+    def __init__(self, config: Phi3Config, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+        elif self.rope_type == "longrope":
+            self._longrope_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+    def _longrope_frequency_update(self, position_ids, device):
+        """Longrope uses long factor if sequence is larger than original pretraining length, short otherwise."""
+        seq_len = torch.max(position_ids) + 1
+        if hasattr(self.config, "original_max_position_embeddings"):
+            original_max_position_embeddings = self.config.original_max_position_embeddings
+        else:
+            original_max_position_embeddings = self.config.max_position_embeddings
+        if seq_len > original_max_position_embeddings:
+            if not hasattr(self, "long_inv_freq"):
+                self.long_inv_freq, _ = self.rope_init_fn(
+                    self.config, device, seq_len=original_max_position_embeddings + 1
+                )
+            self.register_buffer("inv_freq", self.long_inv_freq, persistent=False)
+        else:
+            # This .to() is needed if the model has been moved to a device after being initialized (because
+            # the buffer is automatically moved, but not the original copy)
+            self.original_inv_freq = self.original_inv_freq.to(device)
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+
+
+PHI3_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Phi3Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Phi3 Model outputting raw hidden-states without any specific head on top.",
+    PHI3_START_DOCSTRING,
+)
+class Phi3PreTrainedModel(PreTrainedModel):
+    config_class = Phi3Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Phi3DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+    _supports_attention_backend = True
+    _version = "0.0.5"
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+PHI3_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Phi3 Model outputting raw hidden-states without any specific head on top.",
+    PHI3_START_DOCSTRING,
+)
+class Phi3Model(Phi3PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`]
+
+    Args:
+        config: Phi3Config
+    """
+
+    def __init__(self, config: Phi3Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Phi3RotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
+            )
+            use_cache = False
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        output = BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+        return output if return_dict else output.to_tuple()
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and past_key_values is not None:
+                is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0]
+                if is_padding_right:
+                    raise ValueError(
+                        "You are attempting to perform batched generation with padding_side='right'"
+                        " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
+                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                    )
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if (
+            self.config._attn_implementation == "sdpa"
+            and not (using_static_cache or using_sliding_window_cache)
+            and not output_attentions
+        ):
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                sliding_window=self.config.sliding_window,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        # SlidingWindowCache or StaticCache
+        if using_sliding_window_cache or using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        # DynamicCache or no cache
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+            config=self.config,
+            past_key_values=past_key_values,
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type in ["cuda", "xpu"]
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        config: Phi3Config,
+        past_key_values: Cache,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to plcae the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+            config (`Phi3Config`):
+                The model's configuration class
+            past_key_values (`Cache`):
+                The cache class that is being used currently to generate
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            if config.sliding_window is not None:
+                # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
+                # the check is needed to verify is current checkpoint was trained with sliding window or not
+                if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length:
+                    sliding_attend_mask = torch.arange(target_length, device=device) <= (
+                        cache_position.reshape(-1, 1) - config.sliding_window
+                    )
+                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
+            
+            causal_mask *= diagonal_attend_mask
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                if attention_mask.shape[-1] > target_length:
+                    attention_mask = attention_mask[:, :target_length]
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to(
+                    causal_mask.device
+                )
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+        return causal_mask
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
+class Phi3ForCausalLM(Phi3PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        # self.model = Phi3Model(config)
+        # self.vocab_size = config.vocab_size
+        # self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # # Initialize weights and apply final processing
+        # self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            logits_to_keep (`int` or `torch.Tensor`, *optional*):
+                If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+                If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+                This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Phi3ForCausalLM
+
+        >>> model = Phi3ForCausalLM.from_pretrained("meta-phi3/Phi3-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-phi3/Phi3-2-7b-hf")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past_key_values=None,
+        attention_mask=None,
+        inputs_embeds=None,
+        cache_position=None,
+        position_ids=None,
+        use_cache=True,
+        logits_to_keep=None,
+        **kwargs,
+    ):
+        # Overwritten -- this model may need to switch between short and long rope, invalidating the cache in the
+        # process
+
+        # When the first time input length reached long and short factor switching point, enforce re-compute cache
+        # It will cause downside of slower at this single token position, however, better than current failure.
+        if (
+            past_key_values
+            and self.config.rope_scaling
+            and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1
+        ):
+            past_length = cache_position[0]
+            if past_length <= self.config.original_max_position_embeddings:
+                past_key_values = None
+
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids=input_ids,
+            past_key_values=past_key_values,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            position_ids=position_ids,
+            use_cache=use_cache,
+            logits_to_keep=logits_to_keep,
+            **kwargs,
+        )
+        return model_inputs
+
+
+@add_start_docstrings(
+    """
+    The Phi3 Model transformer with a sequence classification head on top (linear layer).
+
+    [`Phi3ForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    PHI3_START_DOCSTRING,
+)
+class Phi3ForSequenceClassification(Phi3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Phi3Model(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            last_non_pad_token = -1
+        elif input_ids is not None:
+            # To handle both left- and right- padding, we take the rightmost token that is not equal to pad_token_id
+            non_pad_mask = (input_ids != self.config.pad_token_id).to(logits.device, torch.int32)
+            token_indices = torch.arange(input_ids.shape[-1], device=logits.device)
+            last_non_pad_token = (token_indices * non_pad_mask).argmax(-1)
+        else:
+            last_non_pad_token = -1
+            logger.warning_once(
+                f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
+                "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
+            )
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), last_non_pad_token]
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config)
+
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    The Phi3 Model transformer with a token classification head on top (a linear layer on top of the hidden-states
+    output) e.g. for Named-Entity-Recognition (NER) tasks.
+    """,
+    PHI3_START_DOCSTRING,
+)
+class Phi3ForTokenClassification(Phi3PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = Phi3Model(config)
+        if getattr(config, "classifier_dropout", None) is not None:
+            classifier_dropout = config.classifier_dropout
+        elif getattr(config, "hidden_dropout", None) is not None:
+            classifier_dropout = config.hidden_dropout
+        else:
+            classifier_dropout = 0.1
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.score = nn.Linear(config.hidden_size, config.num_labels)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(PHI3_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, TokenClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        sequence_output = self.dropout(sequence_output)
+        logits = self.score(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.config)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )